LLM-Powered Virtual Population for Demand Simulation and Pricing

A new machine-learning framework uses large language models to simulate virtual customer populations, giving retailers a way to forecast demand and set prices for products that have rich descriptions but little sales history, according to research submitted in 2026 [1]. The model, detailed in a paper posted to arXiv, treats potential buyers as random draws from a finite set of customer personas. For each persona, product, and candidate price, an LLM estimates a purchase probability by processing both structured persona attributes and unstructured product information such as text and images [2]. Those persona-level probabilities are then combined through calibrated mixture weights to produce a predictive distribution of aggregate demand [2]. The resulting simulator can evaluate counterfactual prices under different objectives, including expected revenue and risk-aware criteria like conditional value at risk [2]. The authors tested the framework on an online H&M fashion dataset containing product descriptions and images, and report that the calibrated LLM-based simulator achieved the best overall predictive performance among the models examined [2]. The approach is designed for settings where decision-makers need more than a single demand forecast; by generating a full predictive distribution, it allows managers to compare candidate prices, quantify demand uncertainty, and select prices that target either average-case revenue or risk-aware goals [2]. The work arrives as artificial intelligence applications in commerce continue to expand. AI has been used for decision-making, credit scoring, and e-commerce, with recent years bringing major advances in generative models that produce text, images, and other data [5]. The global AI market has seen rapid growth, with India’s AI sector alone projected to reach $8 billion by 2025, expanding at a compound annual growth rate of 40 percent from 2020 to 2025 [3]. While the new demand-simulation framework does not address broader deployment challenges, the paper offers a practical method for using LLMs as demand simulators when historical transaction data is scarce but product information is abundant [2].